Anthropic's IPO Filing Raises the Next AI Question: Who Keeps the Profit?
Revenue is growing exponentially, but the accounting loss and actual cash flow look very different. Growth rate alone will not decide the winners in AI.
$16.3 billion in revenue: first, the comparison period
First-half 2026 revenue of $16.3 billion is 3.5 times full-year 2025 revenue of about $4.6 billion. The periods differ, so the figure should not be read as year-over-year growth. Against first-half 2025 revenue of about $1.2 billion, the increase is far steeper.
Compute and infrastructure costs in the same period were about $8.95 billion, or roughly 55% of revenue. That ratio alone does not give a gross margin. Costs are classified and recognized on different schedules.
- $16.3 billion: first-half 2026 revenue
- $8.95 billion: compute and infrastructure costs, same period
- $693 million: first-half operating loss
- +$3.9 billion: first-half operating cash flow
| Item | Full-year 2025 | First-half 2026 |
|---|---|---|
| Revenue | about $4.6 billion | about $16.3 billion |
| Net loss | about $42 billion | about $80.7 billion |
Periods differ. Net loss includes large non-cash losses on financial instruments.
An $80.7 billion loss is not the same as business performance
Of the roughly $80.7 billion first-half net loss, about $72.4 billion was reported as an accounting charge tied to fair-value changes in financial instruments such as convertible securities. The operating loss for the same period was about $693 million, and operating cash flow was positive at about $3.9 billion.
That does not mean positive operating cash flow is sustainable free cash flow. Working capital swings, prepayments, capital spending and long-term payment commitments all need checking. If convertible securities convert to stock, cash repayment pressure can ease, but existing shareholders may be diluted.
The real AI risk: long-term contracts and falling prices
Anthropic has secured compute capacity over long periods from major cloud providers including Amazon and Google. These companies are both investors and suppliers.
If AI usage keeps expanding, locking in compute early is a competitive edge. If model prices fall faster, or customer demand arrives more slowly than expected, fixed purchase obligations could squeeze profitability. Total contract value should be kept separate from cash already spent and from debt that must be repaid soon.
Who earns the money and who bears the cost
Model companies need to watch customer revenue net of inference costs, plus R&D spending. Cloud companies need to watch rental revenue against capital spending and depreciation. Chip companies need to watch actual build-out speed and reorders.
Investment money and cloud revenue circulating among these companies does not by itself prove economic value for the end customer. The key question is how much of what AI users pay remains at each step of the supply chain.
The listing price is separate from the growth story
A valuation of $965 billion, cited in a private investment in May 2026, is about 29.6 times the simple annualized first-half revenue of $32.6 billion. That is not a formal forward price-to-sales ratio, and it is not an IPO offer price.
The value per share an investor actually gets will depend on the valuation at offering, cash and debt, additional shares issued and the treatment of convertible securities. The conclusion that a company is good should be kept apart from the conclusion that its price is good.
Insight Times Editorial Desk
